Process .opus files with torchaudio

[polinaeterna]

@anton-l suggested to proccess .opus files with torchaudio instead of soundfile as it's faster:

(moreover, I didn't manage to load .opus files with soundfile / librosa locally on any my machine anyway for some reason, even with ffmpeg installed).

For now my current changes work with locally stored file:

# download sample opus file (from MultilingualSpokenWords dataset)
!wget https://huggingface.co/datasets/polinaeterna/test_opus/resolve/main/common_voice_tt_17737010.opus 

from datasets import Dataset, Audio

audio_path = "common_voice_tt_17737010.opus"
dataset = Dataset.from_dict({"audio": [audio_path]}).cast_column("audio", Audio(48000))
dataset[0]
# {'audio': {'path': 'common_voice_tt_17737010.opus',
#   'array': array([ 0.0000000e+00,  0.0000000e+00,  3.0517578e-05, ...,
#          -6.1035156e-05,  6.1035156e-05,  0.0000000e+00], dtype=float32),
#   'sampling_rate': 48000}}

But it doesn't work when loading inside s dataset from bytes (I checked on MultilingualSpokenWords, the PR is a draft now, maybe the bug is somewhere there )

import torchaudio
with open(audio_path, "rb") as b:
    print(torchaudio.load(b))
# RuntimeError: Error loading audio file: failed to open file <in memory buffer>

[lhoestq]

Sounds good to me :) thanks !

[lhoestq]

The path can be None here:

Suggested change

	extension = path.split(".")[-1]
	extension = path.split(".")[-1] if path is not None else None

[lhoestq]

Note that torchaudio is maybe less practical to use for TF or JAX users.
This is not in the scope of this PR, but in the future if we manage to find a way to let the user control the decoding it would be nice

[polinaeterna]

Note that torchaudio is maybe less practical to use for TF or JAX users. This is not in the scope of this PR, but in the future if we manage to find a way to let the user control the decoding it would be nice

@lhoestq so maybe don't do this PR? :) if it doesn't work anyway with an opened file, only with path

[lhoestq]

Yes as discussed offline there seems to be issues with torchaudio on opened files. Feel free to close this PR if it's better to stick with soundfile because of that

[mariosasko]

We should be able to remove torchaudio, which has torch as a hard dependency, soon and use only soundfile for decoding: bastibe/python-soundfile#252 (comment) (opus + mp3 support is on the way).